MISO: Model-Internal-State-Guided Optimization for Ranking Models

2026-08-07

AI Summary

Existing ranking models rely on costly trial-and-error during iterative optimization to adjust architectural components. This work proposes the MISO (Model-Informed Systematic Optimization) workflow, which for the first time systematically leverages internal model states—such as parameters, activations, gradients, and normalization statistics—to generate interpretable editing suggestions. By integrating signal aggregation with adaptive retraining, MISO enables dynamic optimization that strikes a balance between manual tuning and black-box search, substantially improving both efficiency and interpretability. Evaluated on ad ranking tasks, MISO achieves significantly higher normalized entropy while drastically reducing the number of validation runs, outperforming both expert-driven and black-box scaling baselines.

Abstract

Ranking models are repeatedly refined within established model families, yet the choice of which component to scale, replace, or retire is often guided by expensive trial-and-error. We present Model Internal State Optimization (MISO), a systems workflow that uses model internal states (MIS), including parameters, activations, gradients, and normalization statistics, to prioritize such local optimization decisions. MISO extracts MIS from a trained ranking model, aggregates them into ranking, alignment, and comparison signals, and converts those signals into a small set of interpretable candidate edits. Because MIS are re-extracted after each retraining cycle, MISO naturally supports an adaptive optimization workflow that tracks evolving model behavior as data distributions and system requirements shift over time. In an ads ranking case study, MISO improves normalized entropy while requiring substantially fewer validation runs than expert-driven and black-box scaling workflows, offering a practical middle ground between manual tuning and opaque automated search.

Problem

• ranking models
• model optimization
• trial-and-error
• component selection
• adaptive workflow

Innovation

• Model Internal State
• Ranking Models
• Adaptive Optimization
• Interpretable Edits
• Efficient Hyperparameter Tuning